I build software that runs close to the machine: local AI agents, the systems under them, and tools other agents can call.
Yuzu is what I'm shipping now. X · LinkedIn
Yuzu · Swift A local AI agent for macOS. The model runs on the machine through uzu. It connects to Slack, GitHub, Notion, and any MCP server, and learns a team's workflows from skill files. Site
explore-index · Python An agent skill that stops coding agents from re-exploring a codebase they already mapped. Across 31 live runs it cut token use by 39.9% (p = 0.0005) with no loss in accuracy.
Laya · Python A computer-use MCP for macOS. Accessibility builds the table of actions, an MLX model picks one, and the host types it after you approve.
deckdown · JavaScript A Markdown presentation engine for AI agents. Decks render locally to PDF, PNG, and PPTX, with no hosted service in the loop.
impact-compass · TypeScript A CLI that pressure-tests a startup idea against public evidence: demand, competitors, and whether anyone is already paying. No model involved.
Systems work, written from scratch: a 64-bit Rust microkernel, a stack VM and language with its own garbage collector, and a BitTorrent client in Go.
Smaller tools that are still in use: screenshot-mcp captures Linux windows for agents, and wait-whaa-win is a graph-RAG explorer.